Apple Releases LensVLM-9B for Efficient Visual Document Processing

Apple Machine Learning Research has released LensVLM-9B, a 9-billion parameter vision-language model, alongside its research paper, source code, and GGUF quantized model weights. The model introduces a novel selective context expansion technique that scans compressed document images and uses learned tools to uncompress only the relevant pages. Processing long multi-page visual documents usually incurs massive computational overhead because vision-language models require high-resolution image inputs. By scanning low-resolution compressed pages and dynamically zooming into relevant sections, LensVLM drastically improves efficiency without compromising accuracy. LensVLM builds upon modified Qwen model architectures and uses a post-training framework that teaches the model to trigger text expansion for rendered text or high-resolution image expansion for native document layouts. The model weights are provided under the Apple Machine Learning Research Model License, while GGUF versions are available for local deployment.